Multi-Dimensional Tensor Representations for EEG Decoding

20 concepts14 papers1 proposal

Electroencephalogram (EEG) data is inherently multi-modal, simultaneously capturing information across space (where on the scalp neural activity occurs), time (how the signal evolves), and frequency (the underlying neural oscillations). Traditionally, brain-computer interfaces (BCIs) and neural decoding models treat this data as a flat 2D matrix of $\text{Channels} \times \text{Time}$.

However, this traditional representation treats the spatial arrangement of channels as a simple, unordered list, discarding the physical topology of the human head. Two electrodes next to each other in a matrix rows might be on opposite sides of the scalp, while physically adjacent electrodes are separated. To capture the true structure of neural activity, a major research direction focuses on representing EEG data as multi-dimensional tensors (3D or 4D arrays). As noted by Cao et al., the multi-modal nature of EEG makes it naturally suited for tensor representations that preserve and exploit the complex interactions across time, space, and frequency.

Constructing the Spatial-Temporal-Spectral Cube

To transform a standard EEG stream into a multi-dimensional tensor, researchers map the physical coordinates of scalp electrodes onto a structured grid, preserving their spatial relationships.

Traditional 2D Matrix:      3D Tensor Representation:
[Channel 1] ----+            +--------------+
[Channel 2] ----+--> Time    |   Spectral   |
[Channel 3] ----+            |  Bands / Time|
                             +--------------+
                              \  Spatial   \
                               \   Grid     \
                                +------------+

This transformation typically follows one of two structures:

  • Spatial-Temporal Tensors: Electrode signals are mapped onto a 2D grid based on their physical positions on the scalp. These 2D topologies are then stacked along the time axis. For example, Cho et al. interpolate electrode signals into 2D frames to preserve spatial layout, stacking them chronologically to form a 3D spatio-temporal data stream. Yan et al. similarly leverage this 3D spatial-temporal layout to significantly improve auditory attention decoding.
  • Spatial-Spectral Tensors: Instead of raw time series, features like power spectral density or differential entropy are extracted for different frequency bands (e.g., delta, theta, alpha, beta, gamma). Zhong et al. map physical sensor locations onto a 2D grid and stack these grids across multiple frequency bands, creating a structured 3D spatial-frequency matrix. Similarly, An et al. and Chu et al. decompose signals into distinct frequency sub-bands and arrange them into 3D datasets representing fused spatial-spectral features.

For even richer representations, these dimensions can be combined into 4D tensors. Van Dyck et al. and Wu utilize 4D tensors that integrate samples, frequency bands, spatial electrode locations, and time windows into a single unified structure.

Why Tensorization Works: Leveraging Inductive Biases

Structuring EEG data as a tensor works because it aligns the data with the inductive biases of modern machine learning architectures.

Spatial Locality

When electrodes are arranged in a 2D grid that mirrors the scalp, standard convolutional layers can apply local receptive fields to capture spatial correlations, much like processing an RGB image. Zhu et al. demonstrate this principle by transforming EEG data into a multi-channel format analogous to image channels, allowing models to learn directly from the physical proximity of neural signals.

Cross-Frequency and Temporal Coupling

Brain activity is fundamentally broadband, meaning different frequency components are deeply coupled. Van Dyck et al. demonstrate that high-frequency activity (such as low-gamma oscillations) can be predicted with high correlation from lower-frequency bands. A tensor representation allows a model to learn these cross-frequency and cross-electrode dependencies simultaneously, rather than treating each band or channel in isolation.

Preserving Multiway Interactions

Flattening a tensor into a 1D vector forces a model to learn complex multi-linear interactions through dense, fully connected layers, which is highly parameter-inefficient and prone to overfitting. Tensors preserve the multiway structure natively, allowing algorithms to extract clean, non-redundant features.

Algorithmic Frameworks for EEG Tensors

Researchers exploit these multi-dimensional representations using two primary methodologies: unsupervised tensor decompositions and deep neural networks.

Tensor Decompositions

Before passing data to a classifier, unsupervised tensor factorization can be used to isolate underlying neural sources. Ji et al. employ a non-redundant tensor decomposition scheme to capture the complex interactions between different brain signals during simultaneous tasks.

To decode specific cognitive states, Wang et al. apply sparse nonnegative tensor decomposition (NTF) to a four-way tensor (comprising channel, frequency, time, and trial dimensions). This approach enforces sparsity constraints to isolate highly discriminative neural oscillations—such as gamma-band activity—from complex background noise.

3D Convolutional Networks and Transformers

Modern deep learning architectures ingest these multi-dimensional tensors directly for end-to-end decoding. Xu et al. use a multiscale Short-Time Fourier Transform (STFT) to extract features, which are then processed by 3D Convolutional Neural Networks (3D-CNNs) to capture spatio-temporal patterns at various resolutions.

To capture both local spatial features and global contextual dynamics, recent architectures combine convolutions with attention mechanisms. Huo et al. arrange multi-channel frequency data across time and space, feeding the resulting 3D tensor into a hybrid network combining CNNs and Swin Transformers. Similarly, Chu et al. structure EEG data into a 3D matrix representing spatial, temporal, and high-frequency domains up to 75Hz, allowing a specialized Transformer model to capture complex neural patterns that are typically lost in lower-frequency or flattened data formats.

Key Foundational Work

EEG Classification for Hybrid Brain-Computer Interface Using a Tensor Based Multiclass Multimodal Analysis Scheme (2016) — contributed 2 concept elements to this direction.
Single-Trial Decoding of Bistable Perception Based on Sparse Nonnegative Tensor Decomposition (2008) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_3d_eeg_representation_2b4abdee,
  title = {3D EEG representation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/3d-eeg-representation/rdr_2b4abdeee2bd45a8919ff275bf10b6d8},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_3d_eeg_representation_2b4abdee_v1,
  title = {3D EEG representation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/3d-eeg-representation/rdr_2b4abdeee2bd45a8919ff275bf10b6d8/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake